Traffic　forecasting　plays　a　pivotal　role　in　the　advancement　of　intelligent　transportation　systems,　with　significant　implications　for　congestion　alleviation　and　optimal　route　planning.　Existing　approaches　typically　focus　on　capturing　the　temporal　dynamics　of　traffic　states　and　the　spatial　dependencies　across　road　networks　to　improve　prediction　accuracy.　Nevertheless,　two　noteworthy　limitations　persist　in　these　approaches:　(1)　A　lack　of　consideration　for　the　interaction　between　spatiotemporal　features　over　varying　time　scales,　which　impedes　the　effective　utilization　of　traffic　state　information　for　forecasting　future　conditions.　(2)　The　inherent　stochasticity　and　distributional　imbalances　in　traffic　flow,　which　introduce　uncertainty　and　contribute　to　overfitting　issues　in　deep　learning　models.　To　address　these　challenges,　we　propose　a　novel　method,　the　heterogeneous-scale　multi-graph　convolution　networks　based　on　kernel　density　estimation　(KDE-HSMGCN).　This　method　integrates　two　core　components:　the　frequency　feature　layer　and　the　heterogeneous-scale　spatiotemporal　layers.　The　frequency　feature　layer　employs　a　mapping　network　to　learn　and　equalize　traffic　flow　distributions,　mitigating　the　effects　of　distribution　imbalance　and　overfitting　during　model　training.　The　heterogeneous-scale　spatiotemporal　layers　utilize　stacked　spatiotemporal　layers　to　capture　traffic　state　information　across　varying　time　scales.　Experimental　evaluations　on　two　diverse　traffic　datasets　demonstrate　the　superior　performance　of　KDE-HSMGCN　in　medium　and　long-term　forecasting　scenarios.